Features

What it does, and how far it has been proven.

Three labels, and they mean different things. Proven on the rig means a person drove it on a real console with real hardware. Built and measured means audio was measured through it. In development means the code is landing and nothing has proved it yet.

The desk

Proven on the rig

Input channels with head gain, trim, pan, polarity, phantom and pad; a full processing rack per strip; buses, aux sends with pre and post tap points, DCAs, a matrix, mix-minus, stereo channel linking and output inserts.

Sessions persist and reload at boot. Scenes recall. Mute groups, solo (PFL, AFL, solo-in-place with an arming step), a dim, and two panic buttons that separate the mains from the whole console.

Mix-minus is a first-class kind with its own addresses rather than an index-offset hack. Buses and MAIN are strips like any other — there are no master-only mechanisms hiding behind a different code path.

A show load is a clean slate: loading show B over show A returns every channel and every setting B does not mention to what a freshly built strip carries. The one exception is the head-amp, because a show file never asserts phantom, pad or gain onto a real preamp.

Native processing

Built and measured

A parametric EQ bank with a graphical curve, a gate, a compressor, a delay line and a reverb — all of them in the console's own C node, all of them checked against hand-computed values rather than against a recording of themselves.

Moving a fader, a mute, a pan, an EQ band, a gate, a compressor, a delay or a reverb moves audio on the kinds where the console says it does, and where it cannot, it refuses by name rather than lying. That is the strongest part of the console and it is worth saying plainly.

Real-time analyser

Proven on the rig

A native FFT with a multi-resolution zoom tier — bins of about 0.73 Hz below a 300 Hz crossover, where a broad-band analyser tells you nothing useful. Peak hold, pre- and post-EQ overlay, and a per-band difference trace.

Nothing computes unwatched. The analyser is demand-gated per channel and the arming survives a reload because it rides the session, not the browser. Buses have their own analyser, proven by feeding a tone into one bus and reading the spectrum peak at the fed frequency while a bus with no send read a floor with no peak.

The analyser at full height: a live spectrum from 20 Hz to 20 kHz with peak hold and four EQ band handles on the curve.
The analyser running on a live input, with the channel's EQ band handles on the same curve.

Feedback detection and correction

Proven on the rig

The detector watches for the signature of a ringing room and the corrector plants notches into the channel EQ. Two modes with genuinely different behaviour: a ring-out pass plants a true full-cut notch at a very high Q, while live mode plants a gain-carrying bell whose depth comes from the signal's peak-to-average ratio.

Live-mode notches lift again in staged ramps as the room settles — two thirds, one third, gone — and re-clamp with a doubled hold if the frequency comes back. It never touches an operator's own bands; only its own tagged ones.

It is offered on inputs and on every bus strip, aux, mix, matrix and MAIN alike: a single detector on MAIN, planting its notches in MAIN's own EQ, is a sanctioned way to work.

Smart alignment

Proven on the rig

Two microphones on one source arrive at different times. Pick the reference, watch coherence rise, apply. The fit reads the phase slope across frequency, which is why it needs programme material — a pure tone cannot measure delay, because a sine occupies one bin.

It corrects with delay and polarity, and reports the residual rather than silently correcting it. A control that says "137°" without naming the frequency invites the operator to discover that changing the analysis band changes the answer.

A saved show carries the alignment, and loading the show restores it. Loading a show that never had an alignment clears it.

Measured on the rig

Dialled inMeasuredCoherence
0 ms0.000 ms1.000
0.5 ms0.473 ms1.000
2 ms1.903 ms0.998
5 ms4.731 ms0.990
10 ms9.374 ms0.767

Repeated apply converges: 4.748 → 0.252 → 0.016 → 0.001 → 0.000 ms, at a coherence of 1.000.

LV2 plugin inserts

Built and measured

Per-strip and per-output insert chains for LV2 and CLAP plugins, with a plugin picker and an editor generated from the plugin's own ports.

The catalog tiers a plugin by measured round-trip latency — live at 5 ms or under, studio above it, ambiguous ones quarantined — and then ranks within a tier by curation and CPU cost, so the live-safe choice wins by default.

Insert audio is verified by measurement rather than by inspection: a tone through a real graph, a notch racked as an LV2, and a reading at the other end. The reference measurement is a baseline of −12.04 dBFS, the same notch through the plugin door at −100.56 dBFS, and bypass returning to −12.04 exactly — 88.52 dB of cut, through the plugin and not around it.

Plugin delay compensation is applied: the channels that feed one bus are delayed to match the feed with the most insert latency, and on the rig a strip with an out-of-process insert racked measures 0 samples of lag against its pair.

REAC stageboxes

Proven on the rig

Roland S-series stageboxes are discovered, patched and driven — including head-amp control, which is what makes them a stagebox rather than a stream. Two boxes have been established at once on the rig with all 128 links intact across a rebind, and a live 48 to 96 kHz re-pace has been run with the house up.

Phantom power has been confirmed physically at the XLR pins on two inputs with the console independently agreeing. Three sources, one answer — which is the only kind of confirmation a phantom claim is allowed to have.

The transport lives in the sibling FreeREAC projects, not in this project. OpenMixer holds the control path: discovery, head-amp, the patch and the adapter side.

44.1 kHz works as well: a stagebox re-paces to 44.1 kHz and back to 48 kHz with the signal passing through it.

The patchbay: a crosspoint grid of sources against channels, with detected stageboxes grouped down the left edge.
The patchbay. Sources down the side, channels across the top, detected stageboxes grouped and counted.

The web surface

Proven on the rig

A fader wall with banks and user pages, channel strips with a reorderable processing rack, bus-send and matrix views, the patchbay, a graph view, a telemetry view, and a docked multi-window workspace that can be pulled apart and saved as a layout.

The banks roll under the mouse wheel, anywhere over the rail that is not itself a control: one notch is one bank, the same wrapping step the ◀ ▶ buttons make, and a trackpad swipe moves one bank rather than a dozen. A wheel over a fader still trims that fader.

English, Catalan and Spanish. Dark, light and a high-contrast theme, on an axis independent of the console personality.

Its entire state intake is one EventSource and nothing else. The surface cannot open a second channel to the console, by construction rather than by convention.

A change is a PATCH, never a reconnect. There is a deploy gate that drives a real browser through real gestures against a real console and measures long tasks, request bursts, stream re-dials, framework warnings and server-side heals — a green unit suite cannot see a frozen tab.

Control surfaces

Proven on the rig

The Behringer X-Touch and X-Touch Mini are implemented as a bidirectional client: surface gestures become canonical operations, and canonical state drives the motor faders, the LEDs, the V-Pot rings, the scribble strips and the meters. It is wired into the production composition root.

The X-Touch has been run on a real surface: the faders and motors, mute, solo, the V-Pots, SELECT, the master fader, the scribble strips, the meter ladder, a power-cycle and a console restart. The bugs that run found are fixed. The X-Touch Mini and the adapters for Midas, X32/M32 and Roland have not been run on a real surface.

The Digidesign Control|24 and ProControl talk DigiNet over Ethernet. Their protocol library, libdiginet, is released; the console does not drive one. Connecting a Control|24

Recorder and virtual soundcard

Proven on the rig

Multitrack capture to RF64 takes at 48 kHz, a take library, and a virtual soundcard so another application on the same machine can be a source or a destination on the desk.

The code is in: the C capture and playback paths, the RF64 writer and its oracle tests, and the server rows. The arm set is derived rather than stored. The live playback path has no resampler, so a take recorded at another rate is converted offline, on disk, before playback starts.

Recording works: a real take on the rig wrote eighty-four tracks, and a round-trip test records a take and plays it back through the audio graph.

Plugin hosting

The plugin host.

The host that runs the plugins serves the console and Zynthian alike. Each card says where a feature stands on the console, on Zynthian and on any other host. "To confirm" means this state has not been checked on the console yet.

A crash stays in one plugin

If a plugin crashes, only that plugin stops. The rest of your set keeps playing.

Every plugin runs in a worker process of its own, or shares one with the plugins next to it in a chain. A worker that dies is started again by the daemon with its settings and the links it had made, so a bad plugin costs you a moment, not the show.

  • After a killed worker, the daemon put it back with its settings and its links in about 350 ms, on the Pi 5 under Zynthian.Measured with: plugin-hostd 0.19.0, omx-clap-host 0.29.0
  • ConsoleReleasedplugin-hostd 0.16.0
  • ZynthianReleasedzynthian-ui omx-track 2026110701
  • Any hostReleasedplugin-hostd 0.20.0

Detail

A chain in one worker

Plugins in series share one worker and one JACK client. You get most of the memory back.

The daemon plans what can run together: a run of plugins in series is one worker, and a parallel slot, the mixer strip or a jalv plugin ends the run. A processor that runs alone has a Grouping entry in its options that says why. Grouping is off until you turn it on.

  • Memory for 16 plugins in a chain: 55 MB under plugin-hostd, against 99 MB for mod-host and 690 MB for jalv, which runs one process per plugin.Measured with: plugin-hostd 0.18.0, omx-clap-host 0.29.0, mod-host 0.10.9, jalv 1.6.9 · 48 kHz
  • Time to add 16 plugins one by one: 80 ms into a plugin-hostd chain, against 364 ms for mod-host and 13.8 s for jalv.Measured with: plugin-hostd 0.18.0, omx-clap-host 0.29.0, mod-host 0.10.9, jalv 1.6.9 · 48 kHz
  • ConsoleReleasedplugin-hostd 0.16.0
  • ZynthianReleasedzynthian-ui omx-track 2026110701
  • Any hostReleasedplugin-hostd 0.20.0

Detail

A MOD GUI for every plugin

Plugins that ship no interface get one, drawn from the same layout as the controller pages.

omx-panel builds a MOD GUI for each plugin that has none, so MOD-UI pedalboards show a real panel. The same layout feeds the pages on the Zynthian screen, so the two always agree.

  • ConsoleReleasedomx-panel 0.9.1
  • ZynthianReleasedzynthian-sys omx-track 2026110701
  • Any hostReleasedomx-panel 0.12.0

Detail

Ducker, doubler and tone

Music under a voice, a double-tracked sound without a second take, and a tone tilt with a DJ filter.

omx-ducker pulls the programme down while a key is heard, by a set depth with attack, hold and release. omx-doubler adds up to four slightly delayed, detuned and drifting copies spread across the stereo field. omx-tone tilts the whole tone darker or brighter about a pivot and sweeps a resonant DJ filter with one knob.

  • ConsoleComing
  • ZynthianReleasedomx-plugins 0.19.1
  • Any hostReleasedomx-plugins 0.19.1

Detail

Effects that ship with it

EQ, compressor, gate, reverb, delay, drive and a one-load channel strip, as CLAP and LV2.

omx-plugins carries the effects used in the console, each as a CLAP and an LV2 plugin. Add them one by one and choose the order, or load omx strip: trim, filters, a gate, four EQ bands and a compressor in one load and one worker, its five stages in any order.

  • omx strip costs about 5.9 µs per block; omx eq8 and omx comp together about 11.4 µs, with eight EQ bands against four.Measured with: omx-plugins 0.18.0, Raspberry Pi 5 · 48 kHz, 512 frames
  • ConsoleReleasedomx-plugins 0.18.0
  • ZynthianReleasedomx-plugins 0.19.1
  • Any hostReleasedomx-plugins 0.19.1

Detail

What loads is what was measured

A plugin is pinned the first time it is used. If its files change, it is measured again.

The pin ties a plugin name to the exact file that was qualified. The host loads only pinned plugins, so a plugin that was swapped or updated behind its back does not run on trust.

  • ConsoleReleasedplugin-hostd 0.16.0
  • ZynthianReleasedzynthian-sys omx-track 2026110701
  • Any hostReleasedplugin-hostd 0.20.0

Detail

Every plugin measured on your device

You see how heavy a plugin is and whether it is safe before you add it.

plugin-qualify runs each plugin on the device and writes what it found: latency, real-time safety and cost. The menus are built from those records. The records package ships ready-made results for the Pi 5, and a sync in the background checks only the plugins whose files changed.

  • Of 1602 plugins tried on the device, 1079 run in a shared worker and 306 in a worker of their own.Measured with: plugin-qualify 0.18.5, Raspberry Pi 5
  • ConsoleReleasedplugin-qualify 0.18.5
  • ZynthianReleasedzynthian-ui omx-track 2026110701, zynthian-sys omx-track 2026110701
  • Any hostReleasedplugin-qualify 0.20.0, plugin-qualification-records 2026110901

Detail

A plugin that breaks the rules is isolated

A plugin that corrupts the host's own state is moved to a worker of its own.

Every plugin call saves and checks the registers a function must keep. A plugin that changes one is caught and runs alone from then on, so it cannot take the rest of a chain with it.

  • ConsoleReleasedplugin-hostd 0.16.0
  • ZynthianReleasedplugin-hostd 0.20.0
  • Any hostReleasedplugin-hostd 0.20.0

Detail

Big synths use all four cores

A synth that splits its voices into tasks plays more notes, because every core helps.

The host starts helper threads, pinned to the cores the plugin may use. A plugin gets them only after qualification found it safe for real time with the helpers running and faster with them. Any other plugin runs as before. You do not set anything on Zynthian.

  • The mean cost of one block with 48 voices fell from 3335.5 to 857.9 µs, 3.9 times faster, with no block over its deadline. This is a test plugin; Surge XT has no figure of this kind because it does not split its voices.Measured with: omx-clap-host 0.29.0, test plugin of 16 tasks per block, Raspberry Pi 5 · 48 kHz, 256 frames
  • ConsoleReleasedomx-clap-host 0.27.0
  • ZynthianReleasedomx-clap-host 0.31.0
  • Any hostReleasedomx-clap-host 0.31.0, plugin-qualify 0.20.0

Detail

Before you plan a show

Known limitations.

A product page that lists only what works is not a description of a console. These are the limits worth knowing before OpenMixer carries anything you cannot afford to lose.

  • Only the X-Touch has been tested on hardware. The X-Touch Mini and the adapters for Midas, X32/M32 and Roland have not been run on a real surface.
  • The console is not packaged yet.The plugin hosts are: install them from our packages.

This list changes as work lands. The documentation set carries the current version of it.

openmixer

A software mixing console for Linux. The desk is the software; the browser is the surface.

Pages

Licence

OpenMixer is free software under the GPL-3.0-or-later. Every package in the workspace carries the same licence.

Roland, Midas, Behringer, RME and the product names used here belong to their respective owners. OpenMixer is an independent project and is not affiliated with any of them.